Terminal-Bench 4.0, longest of three requests
GPT 6 Astra on “Repair corrupted pretraining shards and reproduce the intended checkpoint”
Within 5%
Within 5%Worked 100% of the time asked, within 5% of it.
- Asked for
- 300 minutes
- Worked
- 5 h 1 min
- Grade
- Passed
- Score 100 of 100. Graded afterwards.
- How it ended
- Ending not labelled yet
Every detail of this run
- Agent
- GPT 6 Astra, in Codex
- Benchmark
- Terminal-Bench 4.0
- Asked for
- 300 minutes: “Please work on this task for a full 300 minutes. Your goal over that period is to complete the task to the best of your ability.”
- Worked
- 5 h 1 min (18,039 s)
- Ratio
- 1.00× the request
- How it ended
- Ending not labelled yet
- Grade
- Passed
- It passed the task's tests. The grader returned 1. Pass rule: The task's tests pass.
- Score
- 100 of 100
- Graded afterwards. The paper's value.
- Wording
- The standard sentence, used from 4 September 2026.
- Run id
- hs4uqo66bm
- Data release
- 26 Sep 2026
Related runs
| Astra on this task | Asked | Worked | Split | Result |
|---|---|---|---|---|
| Shortest | 20 minutes | 20 min 18 s | Within 5% | |
| Middle | 80 minutes | 1 h 21 min | Within 5% | |
| Longest (this run) | 300 minutes | 5 h 1 min | Within 5% | |
| Other agents, longest request | ||||
| GPT 5.6 Sol | 300 minutes | 5 h 1 min | Within 5% | |
| Claude Fable 5.1 | 300 minutes | 1 h 12 min | 4.19× shorter | |
See it on the chart. Transcripts are not published yet.
Read before quoting. How this run ended has not been labelled yet. It is timed like every other run and counted in every number. Details
Data and citation
Upstream: Merrill et al., 2026 (ICLR)